English · 简体中文

LaunchSieve is an Agent Skill for auditing and improving an open-source repository before release. It checks product truth, README and bilingual docs, onboarding, community health, trust signals, accessibility, and visual hierarchy, then proposes the smallest evidence-backed improvements.

The bundled scanner is local and read-only. Repository edits and publishing still require explicit authorization. The Skill ID remains open-source-repo-polish in v0.1.x so existing installations keep working.

Status: v0.1.2 candidate. The bundled audit is local, read-only, zero-telemetry, and dependency-free beyond Node.js 18+.

Why this Skill

Repository polish is not “make a pretty README.” It is the alignment of four surfaces:

Surface	Question
Product truth	Can a visitor understand what the project does and does not do?
Onboarding	Can they reach a verified first result quickly?
Trust	Are evidence limits, security boundaries, maintenance status, and contribution paths explicit?
Visual hierarchy	Does the page guide attention without competing with the project itself?

The Skill combines GitHub repository-health conventions with three original design lenses derived from public guidance:

• Distinctive purpose: choose a specific visual point of view instead of a generic AI template.
• Technical warmth: balance precision, geometry, and human-readable language.
• Calm hierarchy: favor clarity, consistency, accessibility, and restraint over decoration.

It does not copy Anthropic, OpenAI, or Apple logos, fonts, layouts, or protected brand assets.

Quick start

GitHub CLI

GitHub CLI 2.90+ can install the Skill into Codex, Claude Code, Copilot, Cursor, Gemini CLI, OpenCode, and many other supported agents:

gh skill install Offwhite-Del/launch-sieve \
  open-source-repo-polish --agent codex --scope user

Replace codex with claude-code, github-copilot, cursor, or another supported host.

Install as a plugin

ChatGPT Desktop and Codex plugin

codex plugin marketplace add Offwhite-Del/launch-sieve \
  --sparse .agents/plugins \
  --sparse plugins
codex plugin add repo-polish@repo-polish

Start a new task and invoke $open-source-repo-polish.

Claude Code and Claude Code Desktop plugin

claude plugin marketplace add Offwhite-Del/launch-sieve \
  --sparse .claude-plugin plugins
claude plugin install repo-polish@repo-polish

Reload plugins, then invoke /repo-polish:open-source-repo-polish.

Use

Ask the Agent:

Use open-source-repo-polish to audit this repository, propose the smallest
high-value improvements, apply the approved changes, and verify the result.

Or run the deterministic audit directly:

node plugins/repo-polish/skills/open-source-repo-polish/scripts/audit-repo.mjs \
  --root . --json

The script reads only known public repository surfaces. It does not read .env, authentication files, source code, private Agent transcripts, or secret values, and it does not access the network or change files.

Example audit

LaunchSieve 0.1.2
Root: /path/to/repository
Local readiness: 86/100
Findings: 2
- [medium] missing-quick-start: The README has no recognized quick-start heading.
- [low] missing-pr-template: No pull-request template was found.

The score is a navigation aid. Review every finding against the project's maturity and maintenance capacity before adding files.

Safety and privacy

• The script reads only recognized repository documentation and Git status.
• It never reads source code, .env, credentials, private Agent transcripts, or secret values.
• It never accesses the network or changes files.
• Local README targets must remain inside the repository after one URL-decoding pass and symlink resolution; absolute targets are redacted in reports.
• Remote GitHub facts are checked separately by the Agent and must not be inferred from local files.

Workflow

1. Establish repository purpose, audience, current state, and release maturity.
2. Run the local audit and verify GitHub metadata separately.
3. Choose one intentional design direction suited to the project.
4. Separate facts, inferences, proposals, and blockers.
5. Apply the smallest coherent change set on a branch.
6. Render visual assets and validate links, tests, package contents, and secrets.
7. Publish through the repository's normal pull-request workflow.

The Skill stops when all in-scope findings are resolved or have one explicit blocker. A 100% checklist score is not permission to add more decoration.

What the audit checks

• README, license, security, support, contribution, conduct, and changelog surfaces
• Issue and pull-request templates
• Quick start, installation, safety, evidence/demo, and bilingual navigation signals
• Hero/media presence, alt text, badges, and repository-contained local link integrity
• Git branch and dirty-worktree state

The score is a local readiness heuristic, not GitHub's Community Profile score and not a quality guarantee.

Project layout

plugins/repo-polish/skills/open-source-repo-polish/
  SKILL.md                       On-demand workflow
  references/                    Design, health, and evaluation guidance
  scripts/audit-repo.mjs         Read-only deterministic audit
  agents/openai.yaml             Codex UI metadata
.agents/plugins/                 Codex marketplace
.claude-plugin/                  Claude marketplace
test/                            Audit behavior tests

Development

npm test
npm run validate
gh skill publish --dry-run
claude plugin validate plugins/repo-polish

Contributing

See CONTRIBUTING.md, SECURITY.md, and CHANGELOG.md.

License

Apache License 2.0. The linked third-party design guidance remains subject to its own terms and trademarks.